Rocket Quality Evaluation by Computer Vision

Rocket leaves, cultivated in a soilless cultivation system using four different conditions deriving from two distinct irrigation schedules and two different fertilization levels, were put in open polyethylene bags and stored at 10 °C for the number of days required to reach the lowest quality level (QL). At proper times during storage, the samples underwent a sensory evaluation using a 5 level rating scale: 5 = very good (very fresh, no signs of yellowing), 4 = good (fresh, slight signs of yellowing), 3 = fair (slight wilting, moderate signs of yellowing), 2 = poor (severe wilting, evident yellowing), 1 = very poor (unacceptable quality due to decay, severe yellowing and wilting). Level 3 was considered the limit of marketability, while level 2 represented the limit of edibility. Images of rocket leaves, packaged and unpackaged, were acquired by the CVS at each QL and processed to extract colour parameters and to evaluate their QL. The colour features extracted by the CVS were also used to distinguish the irrigation and fertilization managements applied during cultivation. Image processing techniques and machine learning methodologies (including deep learning architectures) were used by the CVS to achieve automatically configuration and settings with minimal human intervention without reducing the effectiveness of results. The same samples underwent physical characterization (colour parameters by the use of a colorimeter, respiration rate and electrolyte leakage) and the results were statistically correlated to colour parameters extracted from images acquired by the CVS to build a model able to assess QLs as well as the corresponding irrigation and fertilization management.

The images acquired by the CVS. Without (on the left) and with (on the right) bag. The colour reference placed in the scene is used to calibrate the colours of each image.